Ain t it Suite? Bundling in the PC Office Software Market. Neil Gandal, Sarit Markovich, and Michael Riordan 1. September 28, 2004

Please Do Not Circulate or Cite. Comments Welcome! Ain t it Suite? Bundling in the PC Office Software Market Neil Gandal, Sarit Markovich, and Michael Riordan 1 September 28, 2004 Abstract Our paper examines the importance of bundling for the evolution of the PC office software market. Using a discrete choice model of product differentiation, our empirical results suggest that word processors and spreadsheets are complements. We also find empirical support for economies of scope for these products. These factors provide incentives for firms to strategically bundle these products. We are especially grateful to Mark Manuszak and thank Jeff Biddle, Chaim Fershtman, Aviv Nevo, Bernie Reddy, Manuel Trajtenberg, Jeff Wooldridge, and seminar participants at Brown University, Carnegie Mellon University, Michigan State University, the University of Pennsylvania (Wharton), Tel Aviv University, and the 2003 IDEI Conference on Software and Internet Industries for helpful comments and suggestions. 1 Gandal: Tel Aviv University, Michigan State University, and CEPR, Markovich: Tel Aviv University, Riordan: Columbia University. 1

1. Introduction There were dramatic structural changes in the office productivity software markets in the 1990 s. The market grew tremendously from 1991-1998, the period for which we have consistent data. There was a shift from DOS based software programs to WINDOWS based software programs. There was also a shift in market leadership away from Lotus (in the spreadsheet market) and Wordperfect (in the word processor market) to Microsoft. Finally, there was a change in marketing strategy from selling components to selling office suites, where the components are integrated together in a single package. The purpose of our paper is to analyze the importance of strategic bundling for the evolution of market structure and the performance of the PC office software market, where office software includes word-processors, spreadsheets, presentation software, database management software, and suites. The literature has identified several incentives for bundling. An incentive for bundling arises when the correlation in consumer values over the products in the bundle is negative. 2 In the case, bundling works like price discrimination to extract consumer surplus. A product complementarity incentive for bundling exists when the products are complements. 3 The incentive for bundling in this case is especially strong when the standalone value of the components is small, relative to the value of the bundle. A third incentive for bundling arises because of economies of scope. When consumer valuations are positively correlated, most consumers are likely to buy both components (if they buy anything). Hence the cost savings of bundling are particularly attractive in the case of positive correlation. 4 Salinger (1995) shows that the economies of scope effect can dominate, 2 See Adams and Yellen (1976), Schmalansee (1982,1984), McAfee, McMillan, and Whinston (1989). The literature goes back to Stigler (1963). He showed that a monopoly can earn additional revenue by either bundling or selling mixed bundles when consumer preferences for the components are negatively correlated. 3 The literature on the bundling of complements began with Cournot (1838). 4 See Salinger (1995). There are, of course, other incentives for bundling as well. For example, entry deterrence might provide an incentive for bundling. In Nalebuff (2004), there is an incumbent that sells in both markets and a potential entrant into one of the two markets. This is not relevant in our setting. 2

so that bundling can be profit maximizing with positive correlation of consumer preferences over the components of the bundle. Casual empiricism suggests that there are likely synergies (complementarities) among the components in computer software office suites because of the links between (and integration of) the components, and because of commands that are common across components. Additionally, consumer preferences for components of office software suites may be positively correlated through an income effect; a high value for a spreadsheet may be due to a high income and this high income would also suggest a high value for a word processor as well. 5 Also, it seems that there would be economies of scope due to common commands across components. We estimate the demand for spreadsheets, word processors and suites. The paper extends the standard discrete choice model of product differentiation to the case when the choices potentially include more than one product. For example, the Microsoft Office suite contains both Microsoft Excel (a spreadsheet) and Microsoft Word (a wordprocessor). Hence, estimation must take account of these combination choices. Previous empirical work has not addressed this issue, despite the fact that this phenomenon is not unique to office software. 6 Our empirical results are consistent with casual empiricism. We find empirical evidence that word processors and spreadsheets are strong complements. We also find some empirical support that consumer preferences over word processors and spreadsheets are positively correlated, although this effect is not statistically significant. Finally, our empirical results suggest that there are economies of scope in marketing and consumer assistance. Our estimates suggest that bundling (rather than selling individual components) is profit maximizing in the case of office suites. Our results hence suggest that economies of scope and 5 Nalebuff (2004) puts forward these arguments. 6 Crawford (2001) empirically examines the importance of bundling in the cable television industry. He shows that the demand for network bundles is more elastic when there are more networks in the bundle. Our approach differs from his in the sense that we allow for, model, and estimate the correlation in unobserved consumer characteristics across products. Gentzkow (2004) employs a methodology similar to ours in examining the online newspaper industry. 3

product complementarities dominate the positive correlation. Salinger s (1995) results -- that economies of scope can make bundling attractive for the firm even when consumers' preferences over the components are positively correlated make our results seem quite plausible. In our case, complementarities among the components contribute as well to making bundling a profitable marketing strategy. Finally, we also find that consumers are willing to pay a significant premium for Microsoft suites, holding the quality of the components constant. Our empirical results are consistent with the notion that only Microsoft successfully integrated the components into a bundle. We then use the estimated parameters to predict oligopoly conduct for several counterfactuals: We first simulate a merger between WordPerfect and Lotus, the dominant firms in the word processing and spreadsheet markets in the DOS era. In the simulation, the merged firm offers a premium suite. We find that these two firms would have earned higher profits in the office software market had they merged and that consumer surplus in the suites market would have been increased by such a merger. We then simulate a market structure in which bundling is not possible (e.g for legal reasons). We find that the attraction of a bundling strategy doesn t depend substantially on whether consumer demands are positively or negatively correlated. This again suggests that our parameter estimates are such that complementarity and economies of scope dominate the positive correlation in consumer preferences. Liebowitz and Margolis (1999) previously studied the evolution of word processor and spreadsheet markets. They argue based on product reviews that Microsoft s dominance of the word processor and spreadsheet markets is due to the superior quality of Microsoft s component products. We conduct an empirical (econometric) analysis and examine how product quality, bundling, and other factors affect demand in the office software market, a market that includes suites, as well as word processors and spreadsheets. Previous empirical work on the software industry has focused on the DOS market and on testing for the presence of network effects. See Gandal (1994) and Gandal, Greenstein, and Salant (1999). 4

The paper proceeds as follows. In section 2, we review the theoretical literature on bundling and discuss the difficulty of theoretically modeling oligopoly competition when firms sell both bundles and component products (mixed bundling). Section 3 discusses the evolution of the PC office software market. Section 4 discusses the data we employ in our empirical analysis. In section 5, we develop the parametric model we estimate and we discuss the estimation algorithm. Section 6 presents the empirical results, while Section 7 uses the estimated parameters to predict oligopoly conduct for counterfactuals. Section 8 briefly concludes. 2. Incentives to Bundle in an Oligopoly The economics literature lacks a well-developed economic theory of bundling by oligopolists selling independent products. There is substantial literature on monopoly bundling (Stigler 1963; Adams and Yellen 1976; Schmalensee 1982, 1984; McAfee, McMillan and Whinston 1989), and an established literature on oligopoly bundling of system components (Matutes and Regibeau 1988; DeNicolo 2000; Nalebuff 2000). There is also a literature on bundling (or tying ) by an incumbent monopolist to deter entry or relax oligopoly competition in a second market (Whinston 1990; Carbajo, demeza and Seidman 1990; Choi and Stefanidis 2001; Naelbuff (2004). But relatively little is known about bundling by multi-product oligopolists selling independent products. There are some results about oligopoly bundling for discrete choice models in which consumer have independent utilities for different products. McAfee, McMillan, and Whinston (1989) extend their theory of monopoly bundling to argue that bundling cannot be absent in the Nash equilibrium in this case. Chen (1997) demonstrates an equilibrium incentive for an oligopolist to differentiate its product by bundling it with a competitively supplied product. But a detailed analysis of equilibrium incentives for bundling in which multi-market oligopolists sell independent products is lacking. A possible model of duopoly bundling is as follows. Assume that Firms 1 and 2 each sell differentiated products A and B. A representative consumer s utility for product i sold by firm f is v if. The four utilities have a joint distribution Fv ( A1, vb 1, va2, v B2) across the population of 5

consumers. Consumers purchase at most one of each product depending on prices. For example, in the case of independent pricing, the firms set prices p if, and a representative consumer is willing to purchase product i from Firm f if and only if v p = max{0, v p, v p } w. The demand for Firm f s product i is the mass of * if if i1 i1 i2 i2 i consumers for which this condition is true. If the firms also offer bundles at prices P f, then the consumers demanding Firm f s bundle are those for whom v + v P = W w + w, * * * Af Bf f max{, A B} where * W va 1 vb 1 P1 vb2 vb2 P2 = max{0, +, + }, and the consumers demanding Firm f s product A alone are those for whom v p = W w w. Demands for the other * * * Af Af max{ B, A} product offers are similar. Given this demand structure, and the costs of producing the products, the two firms simultaneously choose prices ( paf, pbf, P f ) to form a Nash equilibrium. Bundling occurs in equilibrium if paf + pbf > Pf for at least one of the firms. This is a complicated and difficult-to-analyze model, even with strong assumptions on F( v, v, v, v ) and the cost functions. A straightforward approach is to specify A1 B1 A2 B2 F( v, v, v, v ) and a cost function parametrically and analyze the model numerically. A1 B1 A2 B2 Schmalensee (1984) adopted this approach for the monopoly case with some success by assuming a bivariate normal distribution for consumer reservation values and constant returns to scale and scope. Extending this approach to the duopoly case obviously involves a much larger parameter space. We take the more modest approach of estimating a parametric model using data from PC office software market, and analyze numerically the comparative static properties of the estimated model. 3. Evolution of PC Office Software Market, 1991-1998 At the start of the 1990 s, the PC office software market was already well established with a clearly delineated structure. Wordperfect led in the word processor category (Figure 1), Lotus in the spreadsheet category (Figure 2) and presentation graphics, and Borland in database 6

management. These software applications were distinct and sold separately, and overwhelmingly were based on the DOS operating system. The total market for PC office software was approximately $2.6 billion in 1991. The release of WINDOWS 3.0 in 1990, and subsequent improvements, changed all of this. By 1998, Microsoft dominated the PC office software market. The previously distinct applications were bundled in office suites, and overwhelmingly based on the WINDOWS platform. The size of the market had grown to more than $6 billion in 1998. See Figure 3. 1990-1992 was a period of new product introduction and improvement, as competitors adapted to the new WINDOWS platform. Microsoft was first out of the gate with WINDOWS based applications. Microsoft Excel was the first spreadsheet for WINDOWS and Microsoft Office (1990) was the first office suite for WINDOWS. 7 Competitors were later out of the gate, and generally experienced more difficulty ironing out the bugs. Reviews generally agreed that the Microsoft products were superior. Nevertheless, the data clearly show that the switch in platforms from DOS to WINDOWS did not eliminate rivals in the spreadsheet and word processing markets. Lotus acquisition of AmiPro in 1991 enabled it to field a WINDOWS based suite in late 1992. Suites contributed little to industry revenue during this period. The early office suites contained non-integrated word-processor, spreadsheet, database, and graphics programs. Competitors introduced WINDOWS based products later, and generally experienced more difficulties ironing out the bugs. Office suites gathered importance in 1993-94. This was a period of continuous product improvement as office software vendors adapted to an improved version of WINDOWS released in 1992. The new generation of suites were better, but still lacked significant integration. Microsoft was best positioned in the office suite category because it already had highly-rated versions of key underlying components. 7 Samna s Ami (later renamed Ami Pro) was the first word processor for WINDOWS. 7

Microsoft s new office suite, released in early 1994, was extremely well received by computer software trade journals. 8 Microsoft Office 4.2 (including Word 6.0, Excel 5.0 and Powerpoint 4.0) was better integrated than the previous generation of suites and went beyond the standard embedding at the time. Word 6.0 offered a feature where a user could insert an Excel toolbar icon into a document, and then graphically size and place an Excel 5.0 spreadsheet object. 9 PowerPoint 4.0 included a ReportIt feature that took a Presentation and converted it to a Word outline. Microsoft Office 4.2 also included an updated version of Microsoft Office Manager (MOM), a tool that integrated Office applications more tightly. 10 A major reorganization of industry assets followed, as Novell acquired WordPerfect and Borland s QuattroPro in order to field a competitive suite in late 1994. 11 By the end of 1994, WINDOWS dwarfed DOS as a platform for office applications (Figure 4), suites had emerged as the most important product category (Figure 5), and Microsoft had the dominant product in this category (Figure 6). In the summer of 1995 Microsoft released WINDOWS95 and Office 95 simultaneously. 12 Competitors didn't immediately manage to come out with new versions of their own products that took advantage of WINDOWS95. The market for DOS applications all but vanished, and Microsoft s revenue share of the fast growing WINDOWS based office software market surged upward. In 1996, the competition struck back. Corel s Wordperfect Suite and Lotus SmartSuite were well-received and achieved modest market shares (Figure 6). This success led to increased 8 MS Office was awarded the highest overall score by PC/Computing magazine in its February 1994 issue comparing office suites. In the head-to-head comparison, Office outscored all other office suites in each of the five categories, including integration, usability, individual applications, customization and "the basics." Office also swept all the categories in CIO magazine's Readers Choice Awards for Office suites 9 Andrews, Dave It s a Family Affair, BYTE Magazine, 01 November 1993: Vol. 8, No. 12. 10 Nevertheless, Office 4.2 did not offer full integration. Only Excel 5.0 could both control and be controlled by other applications through Visual Basics Applications Edition. Word 6.0 could control another application through VBA but it could only expose its own WordBasic objects so that Excel could use it. PowerPoint 4.0 was not able to control or be controlled by other applications through VBA. 11 The reviewers still weren t persuaded, and Novell eventually exited the industry, selling its office software assets to Corel in 1996. 12 Microsoft announced in July (1995) that it would ship its new version of its popular suite of application programs on August 24 th, the same the day that it intended to release Windows 95. See Microsoft s office suite to be shipped in August, Wall Street Journal, 11 July 1995: Section B5. 8

price competition (see Figure 7), causing revenue growth to slow for the first time. Microsoft Office remained the most highly rated office suite among the three, and by the end of 1998 was dominant in the market. Word Processing and Spreadsheets are by far the most important two components of the PC office software packages Figure 5 shows that these categories are much larger than the Presentation and Database Management Categories. During the 1991-1998 period, word processors, spreadsheets and suites accounted for more than 90% of PC Office software revenue. We focus on these products in the empirical analysis. There were essentially three firms in the office software market: Microsoft, IBM/Lotus 13 and Borland/Corel/Novell/WordPerfect (hereafter Corel/WP). These three firms accounted for at least at least 90% of the WINDOWS office software market from 1993-1998 and 94% of all revenues in every year in the spreadsheet, word processors and suite markets combined during the 1991-1998 period. No other firm had more than a negligible market share in any of these markets during the 1991-1998 period. (See Figure 3.) Hence we limit our econometric analysis to products offered by these three firms. 4. Data Our dataset includes the key office software products: spreadsheets, word processors, and suites. Computer hardware (operating systems) and software are complementary products and the benefit from software consumption can only be realized if consumers have an operating system capable of running the particular software package. In order to focus exclusively on software effects, the sample was restricted to spreadsheets, word processors, and office suites that were compatible with the WINDOWS operating system. 14 Packages that were compatible only with the Apple/Macintosh operating system, for example, were not included. 13 IBM acquired Lotus in 1995. 14 For ease of presentation we refer to WINDOWS for all versions of the WINDOWS operating system made for PCs, including WINDOWS 3.x, WINDOWS95, and WINDOWS98. For the years in which WINDOWS was a graphical user interface that worked with the DOS operating system, we only include products that were made for WINDOWS. 9

Data on prices and quantities (denoted PRICE and QUANTITY) come from two Dataquest/Gartner Reports on Personal Computing Software, one for the 1992-1995 period and one for the 1996-1998 period. 15 Dataquest/Gartner reports (worldwide) shipments and total revenues for each product; hence price is the average transaction price. 16 The variable QUANTITY is the number of units sold (in thousands), and the variable PRICE is the average price. 17 Data on quality of spreadsheets and word processors (denoted QUALITY) come from Liebowitz and Margolis (1999); they employed reviews that gave numerical ratings, and they normalized the top score to 10 in each year. Given the normalization, these scores are not comparable across years. But this is not important, since the choice set is what is available in a particular year. 18 In the case of suites, QUALITY is the sum of the ratings of the relevant spreadsheet and word processor ratings. 19 For example, the rating for the Microsoft Office Suite is the sum of the rating for Microsoft Excel and Microsoft Word in the same year. 20 BUNDLE is a dummy variable that takes on the value one if the consumer purchases either a suite or both a word processor and a spreadsheet. It takes on the value zero if a consumer purchases only a spreadsheet or only a word processor. 15 The first report was purchased from Dataquest/Gartner; we are grateful to Dataquest/Gartner for supplying us the relevant data from the second report. 16 The data on unit sales (or shipments) is comprehensive and includes new licenses, upgrades, and units distributed through original equipment manufacturer (OEM) channels. For the period of our data, office software products were not typically bundled with the operating system. 17 In some cases, we need to average over several versions of the product. For example, in some years, the Microsoft office suite comes in separate versions for WINDOWS and WINDOWS95. There was little difference in price between the versions available for various generations of the WINDOWS operating system. 18 In the case of the LM ratings for Spreadsheets, there are no ratings for 1993 and 1995; fortunately, there are two ratings for 1994 and 1996. We use the first rating in 1994 (which takes place very early in the year) as the rating for 1993; similarly, we use the first rating in 1996 as the rating for 1995. In the case of LM ratings for word processors, there are no ratings for 1996 and 1998. Since there is only a single rating for 1995 and 1997, we average the 1995 and 1997 ratings to obtain ratings for 1996 and use the 1997 ratings for 1998 as well. 19 In theory, it may be preferable to employ a specification where there are two quality variables, one for spreadsheets and one for word processors. However, given the limited number of observations, we want as few parameters as possible. 20 Even if the components had little or no market share, the quality rating for the suite is still the sum of the ratings of the components. Unfortunately, we do not have quantifiable data on the quality of the suite for each year in the sample. 10

We have an additional variable (denoted SCOPE) that measures the scope of the suite in addition to Word Processors and Spreadsheets. Suites also typically include Presentation programs (Powerpoint in the case of Microsoft, Presentations in the case of Corel/WP and Freelance Graphics in the case of IBM/Lotus). For this component, the variable SCOPE gets 1 point. In later years, there is integration with Internet Browsers; in such a case, there is an additional point. For additional components in the Suites (such as email programs, etc.), there is an additional 0.25 points with a maximum of 0.5 points, since we don't want the measure to be affected by under-reporting or over-reporting of minor components. By 1998, all suites obtain the maximum possible score of 2.5. YEARXX is a yearly dummy variable for year 19XX; for example, YEAR93 is a yearly dummy for 1993. YEAR94-YEAR98 are similarly defined. We now define some vendor variables. The variable COREL/WP takes on the value one for Corel/WP word processors and suites, since Word Perfect was the leading word processor before the switch from DOS to WINDOWS. Otherwise, this variable takes on the value zero. It s important to note that this is not a dummy variable for Corel/WP products. This variable measures a reputation effect in the word processing market. Similarly, the variable IBM/LOTUS takes on the value one for IBM/Lotus spreadsheets and suites, since Lotus was the leading spreadsheet before the switch from DOS to WINDOWS. The variable MICROSOFT takes on the value one for Microsoft word processors and spreadsheets, and two for Microsoft suites, since Microsoft emerged as the leader in both of these categories following the shift from DOS to WINDOWS. We have an unbalanced panel of 52 model observations. Microsoft offered all three products in every year. In 1992, there were seven products available since only Microsoft sold suites. 21 In 1993-1995, there were nine products in the sample, as the other two firms offered suites as well. 21 The Lotus/IBM suite was introduced late in 1992 and sales were negligible in that year. 11

In 1996-1998, there were six products available in each year, as IBM/Lotus stopped selling word processors and Corel/WP essentially only sold Suites. 22 Descriptive Statistics are shown in table 1. Variable Mean Std. Dev. Min Max SUITE 0.37 0.49 0 1 SPREADSHEET 0.34 0.48 0 1 WORD PROCESSOR 0.29 0.46 0 1 QUANTITY (000s) 3449 5957 64 32683 QUALITY 23 9.35 0.80 7 10 PRICE ($) 24 126.45 71.77 23.4 350 MICROSOFT 0.77 0.72 0 2 IBM/LOTUS 0.23 0.43 0 1 COREL/WP 0.19 0.40 0 1 SCOPE 1.40 1.89 0 2.5 Y1993 0.17 0.38 0 1 Y1994 0.17 0.38 0 1 Y1995 0.17 0.38 0 1 Y1996 0.12 0.32 0 1 Y1997 0.12 0.32 0 1 Y1998 0.12 0.32 0 1 Table 1: Descriptive Statistics The potential market for office software is defined to be the number of operating systems sold or distributed via OEMs during the relevant year. Our data on operating systems for 1992 comes from Woroch et al (1995), while our data on operating systems for 1993-1998 comes from a Dataquest report on Operating System Shipments. 25 The data in table 2 show that, on average, approximately 80 percent of all consumers with a computer (operating system) purchased an office software product in 1992 and 1993. By 1998, only approximately 50 percent of all consumers purchased an office product. One possible explanation for this decline is that the household market had increased relative to the size of the business market. Indeed, 22 Corel/WP had a negligible share of the word processor market in 1996 and a negligible share of the spreadsheet market during the 1996-1998 period. 23 This statistic does not include the quality of bundles and suites, which is the sum of the qualities of the constituent word processor and spreadsheet. 24 This statistic includes the price of suites, but not the price of mix and match bundles, which is the sum of the prices of the constituent word processor and spreadsheet. 25 The Dataquest reports and the Woroch et al (1995) data delineate between DOS without WINDOWS and DOS with WINDOWS, so it straightforward to simply include the latter. 12

National Telecommunications and Information Administration (NTIA) data show that the percent of households with a personal computer increased in the U.S. from 24.1 percent in 1994 to 36.6 percent in 1997. 26 Year A: WINDOWS Sales of Operating Systems B: Sales of Word Processors C: Sales of Spreadsheets D: Sales of Suites Share of inside goods (B+C+D)/A 1992 11.056 4.650 3.442 0.578 0.784 1993 18.228 6.852 4.640 3.194 0.806 1994 32.107 5.987 5.233 7.689 0.589 1995 54.352 4.693 3.876 12.982 0.397 1996 68.083 2.908 2.979 26.810 0.480 1997 78.406 4.186 2.972 32.977 0.512 1998 89.489 2.091 1.867 38.801 0.478 Table 2: Units of Operating Systems and Office Software Products (millions), 1992-98. 5. Discrete Choice Model and Estimation In this section, we formally specify our discrete choice model. Each consumer compares products across four categories. Consumers can either purchase a spreadsheet only, a word processor only, an office suite, or a mix and match wordprocessor-spreadsheet combination from two different vendors. Hence when all three firms offer word processors, spreadsheets, and office suites, there are 15 possible products : 3 spreadsheets, 3 wordprocessors, 3 office suites, and 6 mix and match wordprocessor and spreadsheet combination from different vendors. 27 Consumers evaluate the products and purchase the one with the highest utility, or make no purchase if that is the best option. The utility from a particular choice is U jk = δ j + θ jk where j indexes the product and k indexes the consumer. Consumer k s utility for choice j has a mean component and a random component that we discuss in turn. The utility from making no 26 See http://www.ntia.doc.gov/ntiahome/net2/presentation/slide14.html. 27 If a product is not available in some year, then the consumer discrete choice problem is modified accordingly. 13

purchase at all is normalized to zero. Optimal consumer choice given these preferences leads to characterization of expected market shares of the products of each vendor. Mean Utility The variable δ j measures the mean utility for product j. We assume that: δ j = β 0 - β 1 * PRICE j + β 2 * QUALITY j + β 3 * YEARXX + β 4 * BUNDLE + β 5 *SCOPE j + β 6 *VENDOR j + ξ j The β ' s are parameters to be estimated. The coefficients β 1 and β 2 capture the effect of the price and quality for a particular software product on the consumer s mean utility, where we assume that the quality of a suite or a mix and match spreadsheet-wordprocessor combination is the sum of the qualities of the individual components. Similarly the price of a mix and match combination is the sum of the prices of the components. The variable BUNDLE measures the complementarity between spreadsheets and word processors. If the estimate of β 4 is greater than (less than) zero, this suggests that the products are complements (substitutes). The variable SCOPE measures the breadth of an office suite and therefore takes on the value zero for word processors, spreadsheets, and mix and match purchases. We also include year dummies (YEARXX) and vendor variables (VENDOR). Note that the coefficient vector is restricted to be the same for all products, and not vary by product category. We do this because, with only a limited amount of data, we are unable to estimate very many parameters with sufficient precision. The variable ξ j is the mean value of any unobserved characteristics of product j. Random Utility 14

The variable θ jk is consumer k s deviation from the mean utility of product j. (The time subscript is suppressed for ease of notation.) Let SPREADSHEET j = 1if product j contains a spreadsheet and SPREADSHEET j = 0 otherwise, and let WORDPROCESSOR j be a similarly defined dummy variable for wordprocessors. Then θ = SPREADSHEET * µ + WORDPROCESSOR * µ + ε jk j 1k j 2k jk The variable µ ik (i = 1 for a spreadsheet and i = 2 for a wordprocessor) is a consumer-specific random utility for a software category. For example, µ 2k > 0 indicates that consumer k has a higher than average value for a wordprocessor. These variables introduce consumer heterogeneity for the demand for different categories of software products. It allows that some consumers place a high value on having a wordprocessor, while other consumers have a great need of a spreadsheet. For suites and mix and match combinations, the consumer receives random utility µ 1k + µ 2k. An important feature of this specification is that it allows a consumer s demand for a wordprocessor to be correlated with the consumer s demand for a spreadsheet. These utility components are assumed to have a symmetric mean-zero bivariate normal distribution, i.e., (µ 1k,µ 2k ) N(0,0,σ 2,σ 2,ρ), where σ 2 is the variance and ρ is the correlation coefficient. We estimate the parameters of this distribution, with a particular interest in the correlation coefficient. We restrict the relationship between variance and correlation coefficient, as explained below when we discuss the estimation algorithm. ε jk is consumer k s additional random utility for product j. This term introduces an additional source of consumer heterogeneity, i.e. some consumers may be more attracted to a particular product. Unobserved consumer heterogeneity in preferences over products in a particular software category or products involving two software categories enters only through this variable. The ε jk are assumed to be independently and identically distributed according to a standard double exponential distribution. This captures an idiosyncratic preference for individual products. This is the error structure employed in the usual logit demand model. It permits a convenient characterization of expected market shares, as described below. 15

Market shares Given the logit structure of demand derived from the distributional assumptions on ε jk probability that consumer k chooses product j conditional on ( (, ) µ µ ) is 1k 2k, the P e = + = δ µ µ j+ SPREADSHEET j* 1k+ WORDPROCESSOR j* 2k jk 15 δ l+ SPREADSHEETl* µ 1l+ WORDPROCESSORl* µ 2l 1 e l 1, and the probability that consumer k makes no purchase is P = + = 0k 15 l SPREADSHEETl* 1l WORDPROCESSORl* 2l 1 e δ + µ + µ l 1 1. These probabilities can be employed in a straightforward way to simulate market shares for office suites. The calculations for an individual software category are somewhat more complicated. Consider for example a particular vendor s wordprocessor. Let product j refer to the standalone wordprocessor, and let j and j refer to the two mix and match combinations that involve that wordprocessor. Then the probability that consumer k purchases this vendor s word processor (separately from the suite) is Pjk ' + Pj'' k+ Pj''' k. Making similar calculations for the wordprocessors of other vendor s, it is straightforward to calculate simulated market shares in the wordprocessor category. Obviously, the validity of these calculations requires a large number of consumers. Estimation Algorithm Let µ 1k =Y 1k + αy 2k and µ 2k =αy 1k + Y 2k where Y 1k and Y 2k are i.i.d. standard normal random variables. Then (µ 1k, µ 2k ) N(0,Σ) with 16

2 1+ α 2α Σ= 2 2α 1+ α 2α The correlation coefficient is ρ =. Thus there are two values of α (and hence two values 2 1 + α of ρ) that yield the same value of σ 2, one positive and one negative. The estimation algorithm proceeds in several steps. Step 1: Take random draws of (, ) makes a single choice. 28 X X for 100,000 consumers per year. Each consumer 1k 2k Step 2: Assume an initial value for α, and find δ using the contraction mapping δ j,new = δ j,old + ln( actual market shares) ln( simulated market shares) until convergence is obtained. 29,30 Step 3: Given δ, we compute the implied values of the unobservables, i.e., ˆ ξ = ˆ δ X ˆ β, where X is the matrix of right hand side variables, Z is the matrix of exogenous right hand side variables and instrumental variables, 31 β ( X ZWZ X ) 1 estimated parameters, and the weighting matrix W=(Z Z) -1. 32 ˆ = ' ' X ' ZWZ ' ˆ δ is the vector of 28 We abstract from the issue of repurchases and upgrades. 29 The initial value of δ j comes from δj = ln(sj)-ln(s o ), where s o is the share of the outside good. See Berry, Levinsohn, and Pakes (1995) for details. 30 Since the data consist of sales of spreadsheets, wordprocessors and suites, the 15 choices are mapped into the 9 products. This is straightforward (as described above) since the total number of Microsoft Word wordprocessor sales (separate from the suite) is the number of consumers who purchased Word as a standalone product plus the number of consumers that mix and match, i.e., those that purchased Word with The Lotus/IBM spreadsheet and Word with the Corel/WP spreadsheet. 31 Since price is endogenous, we instrument for it. Following other authors, we use the average software quality of other products in the same year and category (word processor, spreadsheet, and suites) as an instrument for price. In years, when there are no other competing products in the category, the quality of the other products is zero. 32 As Nevo (1998) notes, this weighting matrix yields efficient estimates under the assumption that errors are homoskedastic. 17

Step 4: We evaluate the GMM objective function ˆ ξ ' ZWZ ' ˆ ξ Step 5: We then update the guess for α, and return to step 2. 33 Since price is endogenous, we instrument for it. Following other authors, we use the average quality of characteristics of other products as instruments. In particular, we have three instruments for price: (I) the average quality of other products sold by the same firm in the same year, (II) the average quality of other products in the category (word processor, spreadsheet, and suites) in the same year, and (III) the average scope of all other products in the same year. In years, when there are no other competing products in the category, the quality of the other products is zero. 6. Empirical Results Table 3 contains our estimates when we estimate the demand equation alone. Variable Coefficient T-Statistic Coefficient T-Statistic Demand Constant -2.54-1.54-0.20-0.64 Price 0.041 1.10 0.070 0.91 Quality 0.31 0.86 0.51 0.82 Scope 0.45 0.93 0.067 0.047 Bundle 3.13 2.40 5.02 1.73 Microsoft 4.39 1.73 5.77 1.32 Corel/WP 3.63 1.43 5.45 1.07 IBM/LOTUS 3.66 1.32 5.80 1.02 Alpha (α) 34 0.30 0.16 0.27 0.02 YEAR93-0.85-0.84-1.86-0.76 YEAR94-1.88-1.39-3.43-0.95 YEAR95-2.95-1.96-5.18-1.00 YEAR96-3.93-1.36-7.30-0.91 YEAR97-4.18-1.41-8.61-0.92 YEAR98-4.39-1.52-8.61-0.92 Installed Base/10000 4.96 0.62 GMM objective 3.44 2.92 Table 3: Demand Equation Estimates; Instruments Quality & Scope of other brands 33 The estimate of α is updated by Matlab using the Golden Section search and parabolic interpolation. 34 This implies a correlation coefficient, ρ=0.55. The standard error of ρ can be found by the delta method. 18

The difference between the first and second set of estimates in Table 3 is that in the second regression we have added Installed Base, a variable that captures the benefit associated with a larger network. 35 This benefit would primarily be due to the ability of users to share files. Since the various brands are essentially compatible for the period of our data -- for example, word processing documents written in Word Perfect could be read into Microsoft Word and edited -- there would not be an advantage to having a larger installed base one. The second regression indeed shows that Installed Base has the expected sign, but is statistically insignificant, that is, there is not any evidence of any network effect advantage. This makes sense given the discussion above. Table 3 also shows that the addition of installed base does not qualitatively change the results, that is, all coefficients have the expected signs in both regressions in Table 3. However, the addition of installed base increases the standard errors of the regression and that the t-values are lower in the second regression. We prefer the first regression in Table 3. The first regression in Table 3 shows that the estimated coefficients on PRICE and QUALITY are not statistically significant. The coefficient on BUNDLE is positive and statistically significant, suggesting that word processors and spreadsheets are indeed complementary products. The coefficient estimate associated with the variable MICROSOFT is statistically significant at the 90% level of significance. Although the coefficients on IBM/ LOTUS and COREL/WP are not statistically significant, they are roughly of the same magnitude as the coefficient on MICROSOFT. Recall the definitions of these variables, this suggests that while IBM/Lotus provided significant competition in the spreadsheet category and Corel/WP provided strong competition in the word processor category, neither firm provided significant competition in the suite category. That is, the IBM/Lotus and Corel/WP suites were essentially viewed as offering little more than the relevant component for which the firm was successful in the DOS market. Since we have controlled for the quality of the components, the results are consistent with the notion that only Microsoft successfully integrated the components into a bundle. 35 Explain how we calculated installed base and that it really doesn t matter for the estimates. 19

Casual empiricism indeed suggests that the other suites were not integrated as well as Microsoft s suite. 36 Liebowitz and Margolis (1999) note, When they [Microsoft s competitors] did assemble competing suites, they tended to cobble together products that had little in common. Stan Miastkowski, writes about the 1997 Corel/WP as follows: Prior versions of WordPerfect Suite showed the results of cobbling together a bunch of disparate applications 37 Data compiled from trade journals, as summarized in Appendix B, are consistent with the above assessments. 38 Although the estimated coefficient on SCOPE has the expected sign, it is not statistically significant. The yearly dummy variables capture shifts in the difference between the value of office software products and the outside option. The coefficients associated with the yearly dummies are declining in value. This is in large part due to the fact that the consumer purchases of spreadsheets, word processors and suites divided by the number of operating systems was declining as well. That is more consumers elected not to purchase an office software product in later years. Indeed the correlation between the percentage of consumers choosing to purchase an office software product and the coefficients of the yearly dummy variables is 0.88. The estimated value of α implies a correlation coefficient of 0.55. This suggests a positive correlation in preferences for word processors and spreadsheets, the two most important components of the office software market. However, the standard error of the estimate of α (and hence ρ as well) is quite large, leading to a statistically insignificant estimate. In order to examine this effect further, we re-estimated the equation for different (fixed) values of α (ρ). The estimated coefficients do not change very much as ρ changes, yet the GMM objective 36 It is not possible to formally add integration or cross application compatibility to the data set since (as noted above) we do not have data on these measures for each year to year. 37 See Corel s Nearly Perfect Suite Spot, Byte.com, July 1997, available at http://www.byte.com/art/9707/sec11/art4.htm#077ev2t1 (accessed September 29, 2004). 38 Indeed, it seems that Microsoft still retains a significant advantage in its ability to integrate the components better than its competitors. ZDNet reviews of the most recent versions of suites, as summarized in Appendix B, suggest that the Corel/WP and IBM/Lotus suites still do not integrate the components of their suites as well as Microsoft does. 20

function is quite flat as a function of ρ. This suggests that although it is not easy to estimate ρ with a great deal of confidence, the estimates of the other coefficients are quite robust to changes in ρ. In Table 4 we estimate the both the demand and oligopoly pricing (supply) side of the market under the assumption that the firms compete on prices. A Nash equilibrium can be estimated by employing the first order conditions. It can easily be shown that the first order condition for firm j is p j =1/[β 1 (1-s j )] + mc j, where mc j is the marginal cost of product j and s j is the market share of product j. Letting mc j = w j γ + υ j, where the w j represent the variables that affect marginal costs, the supply equation to estimate is 39 p j = 1/[β 1 (1-s j )] + w j γ + υ j We include the following cost variables in the supply side: (I) Yearly time dummies, (II) product category dummy variables (Wordprocessor and Suite) and (III) firm dummy variables (MS and IBM/LOTUS). 40 The results from jointly estimating both the demand and oligopoly sides of the market are shown in Table 4. 39 For simplicity, we assume that the firms do not take into account the elasticity of demand among their products. 40 Note that these are not the same as the demand side vendor variables. 21

Demand Supply Variable Coefficient T-Stat Variable Coefficient T-Stat β 1 (Both equations) 0.072 1.77 β 1 0.072 1.77 Constant 0.23 0.15 Constant 134.11 97.34 Quality 0.49 1.29 Scope 0.37 0.39 Bundle 2.04 1.49 Microsoft 6.55 2.39 MS Dummy 69.44 Corel/WP 5.83 2.15 IBM/LOTUS 5.98 2.02 IBM dummy 9.72 Alpha (α) 41 0.19 0.02 YEAR93-1.95-1.70 YEAR93-30.65 YEAR94-3.90-2.58 YEAR94-48.75 YEAR95-5.64-3.38 YEAR95-68.73 YEAR96-7.78-2.40 YEAR96-103.35 YEAR97-8.12-2.48 YEAR97-105.71 YEAR98-8.34-2.59 YEAR98-104.92 Wordproccessor 13.19 Suite 28.53 GMM Objective 56.99 Table 4: Parameter Estimates and T- Statistics: Supply and Demand Table 4 shows that the demand side parameter estimates are qualitatively unchanged. The increased efficiency associated with the full information estimation method leads to several more statistically significant coefficients. In particular, the coefficient on PRICE (β 1 ) is statistically significant at the 90% level and the coefficients on the vendor dummies are all significant at the 95% level of confidence. 42 From the supply side, we can see that there are economies of scope, especially for Microsoft. Estimated marginal costs for 1998 are as follows: (The estimated 1995 costs are approximately $55 more for each category; in 1998, the IBM/Lotus and Corel/WP only sold suites.) 41 This implies a correlation coefficient, ρ=0.37. The standard error of ρ can be found by the delta method. 42 The coefficient on QUALITY is more significant as well, although it is not statically significant. 22

Firm Spreadsheet Word processor Suite Microsoft $90 103 120 IBM/LOTUS 30 43 60 Corel/WP 20 33 50 Table 5: Estimated Marginal Costs As discussed in the introduction, economies of scope make bundling more attractive. This is especially true when (as we found) consumer preferences are positively correlated. 6.1 The Current Population Survey In order to further assess whether our estimate of the positive correlation and complementarity is reasonable, we obtained survey data from the Current Population Survey (CPS) Supplement on Computer and Internet use from September 2001. 43 The supplemental data on computer and Internet use were first collected in 1998. However, questions about spreadsheet and word processor usage were only asked beginning in 2001. There were approximately 180,000 individuals in the 2001 CPS Supplement. The CPS uses weights to produce basic demographic and labor force estimates. In 2001 the following questions were asked about spreadsheet and word processors for both home and office use: 44 Do you use the computer at home (at the office) for word processing or desktop publishing? Do you use the computer at home (at the office) for spreadsheets or databases? The weighted results are shown in the following tables. 43 The CPS is a joint project of the Bureau of Labor Statistics and the Bureau of the Census. See http://www.bls.census.gov/cps/ for more details. 44 The possible answers are either yes or no. 23

Home Use Use Spreadsheets? Use WPs? Yes No Yes 0.27 0.32 No 0.05 0.36 Office Use Use Spreadsheets? Use WPs? Yes No Yes 0.50 0.17 No 0.12 0.21 Table 6: CPS Supplement on Computer and Internet In the case of home (office) use, 61% (71%) of the individuals answered either yes to both of the questions or no to both of the questions. This provides some support for positive correlation and/or complementarity. 7. Counterfactuals This section uses the demand side estimates in Table 4 to predict oligopoly conduct for counterfactuals. Prices Costs Consumer share Profits Microsoft Lotus WordPerfect Microsoft Lotus WordPerfect Microsoft Lotus WordPerfect Microsoft Lotus WordPerfect (1) (2) (3) (4) 158.97 158.44 157.69 109.03 124.10 109.04 124.11 110 110 85 85 27.4 27.3 0.06 0.06 13.42 13.22 0.15 0.15 110 85 85 27.4 0.7 0.8 13.08 0.16 0.16 Table 7: Competitive Effects of Bundling (QUALITY=10, ξ=0) 151.21 138.53 110 85 26.4 10.4 10.90 2.96 Assumptions: 1. Microsoft is a monopolist selling Office suite 2. Microsoft suite competes with Lotus spreadsheet and WordPerfect word processor 3. Microsoft, Lotus, and WordPerfect compete in the market for suites 4. Microsoft competes against merged Lotus-WordPerfect suite. 45 45 In order to conduct the counterfactual, in the case of the variable QUALITY, we employ the sum of the Corel/WP wordprocessor and the IBM/Lotus spreadsheet. Since the scope of both the Corel/WP and IBM/Lotus suites were identical in 1998, we use that value. 24

The above table shows that Microsoft s predicted prices and profits are essentially the same regardless of whether it (I) has no competition in the suite market, (II) competes with the Lotus spreadsheet and the WordPerfect word processor, or (III) competes against the Lotus and Word Perfect suites. This is not surprising given our estimates and previous discussion. The interesting case is when Lotus and WordPerfect merge and offer a high quality suite. The above table shows that many more consumers purchase new combined suite and that this effect is primarily due to an increase in the number of consumers who purchase suites rather than the outside good. Nevertheless, the increased competition leads to a predicted five percent decline in Microsoft prices. Further, its (inside goods) share of the suite market falls from 95 percent to 72 percent. The combined Lotus and WordPerfect suite charges a higher price due to the improved quality. Despite the higher price, the size of the market grows by 27 percent. SCOPE=0 1998 Monopoly Pure Independent Pricing Bundling Price 162.26 167.91 158.49 129.76 106.22 82.44 Consumer Share 36.6 36.6 26.3 2.9 9.3 25.6 Unit Cost 110 110 110 110 85 55 Profit per consumer 19.24 21.08 12.66 1.13 3.93 14.05 Table 8: Microsoft Suite Prices and Profits (ρ=0.5) SCOPE=0 1998 Monopoly Pure Independent Pricing Bundling Price 162.26 158.29 149.29 131.65 105.81 79.75 Consumer Share 36.6 40.4 27.2 2.1 7.4 25.3 Unit Cost 110 110 110 110 85 55 Profit per consumer 19.24 19.67 10.60 0.907 3.07 12.50 Table 9: Microsoft Suite Prices and Profits (ρ=-0.5) 25